System and Method for Mental Diagnosis Using EEG

The present patent application claims priority to U.S. Provisional Patent Application No. 63/349,296, filed Jun. 6, 2022, and entitled “System and Method for Mental Diagnosis using EEG”, the disclosure of which is incorporated herein by reference thereto.

Mental disorders are very costly not only for the affected individual and their family but also for the society as a whole. The World Health Organization estimates that over 1 billion people worldwide suffer from some form of mental disorder. Further, by the year 2030, over 6 trillion dollars may be spent on treating the mental disorders. A further 16 billion dollars may be lost through an estimated 12 billion days of work lost each year due to the burden of mental disorder.

In the Unites States of America alone, approximately one in five adults experiences a mental disorder in a given year, 18.1% of adults experience an anxiety disorder, such as posttraumatic stress disorder, obsessive-compulsive disorder and specific phobias, 6.9% of adults have at least one major depressive episode each year, and 1.1% of adults live with schizophrenia.

Major psychiatric disorders, including major depressive disorder (MDD), bipolar disorder (BP), and schizophrenia (SZ), are among most severe and impactful mental illnesses. They often have strong comorbidities and share substantial features. Approximately 85%-90% of patients with depression also experience symptoms of anxiety, and vice versa. The simultaneous presence of two or more psychiatric diseases are associated with greater severity, worse response to the pharmacological treatment and have a greater risk of suicide than either condition alone.

The consequences of the lack of treatment for these mental disorders are significant. In the United States of America, mental disorders are the third most common cause of hospitalization for both youth and adults aged 18-44. Suicide is the tenth leading cause of death, and the second leading cause of death for those aged 15-24. And each day, approximately 18-22 Armed Forces veterans die by suicide.

A key factor in treatment of mental disorders is proper diagnosis. Typically, the standard method of diagnosing mental disorders includes either the use of the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, American Psychiatric Association, (2013), Arlington, Va. (“DSM” of “DSM-5”) or the International Statistical Classification of Diseases and Related Health Problems (ICD), Chapter 5: Mental and behavioral disorders, 10th Revision (ICD-10), 1994, Geneva: World Health Organization (“ICD”). Both of these standards primarily involve diagnosis using conversation with the patient regarding symptoms and behavior.

Therefore, various other diagnostic approaches employ identification of disease-related objective biomarkers or subtypes obtained using structural magnetic resonance imaging (sMRI), or functional connectivity features derived from resting-state functional MRI. For example, by analyzing canonical correlations between resting-state fMRI functional connectivity and Hamilton Depression Rating Scale (Ham-D or HAMD), biomarkers were extracted for MDD subtype identification and predicted transcranial magnetic stimulation (TMS) treatment outcomes.

At present, psychiatric diagnoses are based entirely on phenomenological descriptions, with no assay-based biological criteria underlying diagnosis. This has resulted in syndrome categories within serious mental illness that have overlapping symptom fields. The disadvantage of the diagnostic evaluations being only subjective—that is, based on the interviewer's own perceptions—may lower the diagnostic reliability and the resultant treatment and, sometimes, may result in two clinicians forming two different diagnoses of the same patient. For example, due to the overlapping clinical symptoms at onset, about 60% of BP patients are initially misdiagnosed as MDD and have to wait 5-10 years before receiving appropriate diagnoses.

In addition, the biological heterogeneity of psychiatric disorders has a substantial effect on treatment outcome as well, resulting in the unpredictability of therapeutic effects based on pretreatment clinical symptoms. For example, the treatment response of treatment-resistant major depressive disorder using transcranial magnetic stimulation (TMS) varies from 45% to 60%. Thus, obtaining the correct mental disorder diagnosis and then accurately predicting the TMS-treatment response is imperative to best support recovery of the patient. As for TMS treatment, currently, the treatment plan of the TMS is designed entirely based on pretreatment clinical symptoms, rather than objective biomarkers.

Therefore, an accurate diagnosis based on objective neuroimaging biomarkers, which could benefit from machine learning methods, is important for making an individualized treatment plan. However, the steps of acquiring these MRI neuroimages are complex, time consuming and expensive.

Accordingly, a need arises for systems and methods to easily and accurately predict a mental disorder and provide to the user a proper TMS-treatment response predictor.

Aspects of the disclosure relate to systems and methods for diagnosing mental illness in a patient. An exemplary method for diagnosing mental illness in a patient may comprise the steps of measuring a plurality of encephalography signals from a patient. The method may further comprise feeding the plurality of encephalography signals into a trained multi-scale recurrent neural network which may classify the patient as having a selected mental illness of a plurality of mental illnesses. Then the method may communicate the selected mental illness to a user.

In an embodiment, the plurality of encephalography signals comprises at least one electroencephalography (EEG) signal or at least one magnetoencephalography (MEG) signal. The EEG signal may comprise at least one of the bilateral front (Fp1, F8) channels or the parietal-occipital (P7, P3, P8, PO4, O1, O2) channels. The plurality of mental illnesses may comprise major depressive disorder, bipolar disorder, schizophrenia, and a normal control. The method may further comprise extracting a feature from the multi-scale recurrent neural network and applying the feature to a second machine learning model to classify a sub-type of the selected mental illness. The method may further comprise incorporating other patient information as part of signals input into the trained multi-scale recurrent neural network.

In an embodiment, the method may further comprise, if the patient is diagnosed as having a major depressive disorder, reducing the dimensionality of the plurality of encephalography signals and visualizing the dimensionality-reduced plurality of encephalography signals. The method may further comprise classifying the patient as belonging to a sub-type of major depressive disorder, based on the visualization, and communicating the sub-type to the user. In an example, the sub-type may comprise a major depressive disorder treatable with transcranial magnetic stimulation. A method of dimensionality reduction and visualization may be t-distributed stochastic neighbor embedding (t-SNE). A sub-type may also be identified by the use of a clustering technique.

In an embodiment, this disclosure relates to a system for diagnosing a patient has having a mental illness. The system may comprise a computer with memory and a processor, a user device, and a measurement device. The system may measure, using the measurement device, a plurality of encephalography signals from a patient. The system may then feed, at the computer, the plurality of encephalography signals into a trained multi-scale recurrent neural network, which classifies the patient as having a selected mental illness of a plurality of mental illnesses. The system may then communicate the selected mental illness to the user device.

In an embodiment, the plurality of encephalography signals comprises at least one electroencephalography (EEG) signal or at least one magnetoencephalography (MEG) signal. The EEG signal may comprise at least one of the bilateral front (Fp1, F8) channels or the parietal-occipital (P7, P3, P8, PO4, O1, O2) channels. The plurality of mental illnesses may comprise major depressive disorder, bipolar disorder, schizophrenia, and a normal control. The computer may further perform the steps of extracting a feature from the multi-scale recurrent neural network and applying the feature to a second machine learning model to classify a sub-type of the selected mental illness. The computer may also perform the step of incorporating other patient information as part of signals input into the trained multi-scale recurrent neural network.

In an embodiment, if the patient is diagnosed as having a major depressive disorder, the computer may further perform the steps of reducing the dimensionality of the plurality of encephalography signals and visualizing the dimensionality-reduced plurality of encephalography signals. The computer may further perform the steps of classifying the patient as belonging to a sub-type of major depressive disorder, based on the visualization, and communicating the sub-type to the user. In an example, the sub-type may comprise a major depressive disorder treatable with transcranial magnetic stimulation. In an example, the dimensionality reduction and visualization may comprise t-distributed stochastic neighbor embedding (t-SNE). The sub-type may also be identified by the use of a clustering technique.

Other features of the present embodiments will be apparent from the Detailed Description that follows.

In the following detailed description of the preferred embodiments, reference is made to the accompanying drawings, which form a part hereof, and within which are shown by way of illustration specific embodiments by which the invention may be practiced. It is to be understood that other embodiments may be utilized and structural changes may be made without departing from the scope of the invention. Electrical, mechanical, logical, and structural changes may be made to the embodiments without departing from the spirit and scope of the present teachings. The following detailed description is therefore not to be taken in a limiting sense, and the scope of the present disclosure is defined by the appended claims and their equivalents.

The present disclosure relates to diagnosing major psychiatric disorders including major depressive disorder (MDD), bipolar disorder (BP), and schizophrenia (SZ) as distinct from normal controls (NC). These disorders are among the most severe and impactful mental illnesses. They lead to decreased quality of life, premature death and disability in many patients, and increased health care costs. Major psychiatric disorders often have strong comorbidities and shared substantial features, causing difficulties in diagnosis and treatment. The current psychiatric diagnosis guidelines are based on phenomenological descriptions, with no assay-based biological criteria underlying diagnosis. This has resulted in syndrome categories within major psychiatric disorders that have overlapping symptom fields. For instance, due to the overlapping clinical symptoms at onset, about 60% of BP patients are initially misdiagnosed as MDD and have to wait 5-10 years before receiving appropriate diagnoses. The biological heterogeneity of psychiatric disorders has a substantial effect on treatment outcomes, resulting in the unpredictability of therapeutic effects. For example, the treatment response of treatment-resistant major depressive disorder using transcranial magnetic stimulation (TMS) varies from 45% to 60%. By defining clinical-relevant “disease subtypes” based on biological biomarkers, more accurate therapy plans are promised to be made according to the characteristics of each subtype. Therefore, studying the heterogeneity and subtypes of psychiatric disorders based on objective biomarkers is urgent for accurate diagnosis and treatment.

Recent studies have adopted various approaches that transcend traditional diagnostic boundaries for identifying disease-related biomarkers or subtypes by using structural magnetic resonance imaging (sMRI), functional connectivity features derived from resting-state functional MRI, or resting-state electroencephalography (rsEEG). For example, two clinically relevant subtypes of post-traumatic stress disorder and MDD were identified based on functional connectivity patterns in EEG. By analyzing canonical correlations between resting-state fMRI functional connectivity and the Hamilton Depression Rating Scale, researchers extracted biomarkers and identified MDD subtypes of which TMS treatment outcomes could be predicted. (Drysdale, A. T. et aL. “Resting-state connectivity biomarkers define neurophysiological subtypes of depression.”23, 28-38, (2017).) Compared to MRI, EEG is an inexpensive measurement with higher time resolution and non-magnetic effects, making it a practical healthcare tool in a variety of clinical environments. Decades of studies on EEG have provided us with multiple methods (e.g., coherence, phase synchronization, phase-slope index, Granger causality) to quantify neural interactions as well as to provide a valid interpretation of the findings. However, due to the respective assumptions, the EEG metrics have limitations in fully leveraging the spatiotemporal information in EEG. Besides, the choice of a variety of hyperparameters such as frequency band makes the metrics challenging for subtype discovery. Moreover, non-linear approaches which are promising in extracting discriminative information from the time domain of rsEEG have not been well studied in the literature. Compared to the standard machine learning method, deep learning can encode more robust discriminative neuroimaging representations by characterizing potentially non-linear high-level patterns existing in the input features. The convolutional recurrent neural network, a specific deep learning architecture, has been proved efficient in mental disorder classification tasks by leveraging spatiotemporal information from fMRI time sequences. Different from the popular frequency-based EEG methods as reported by others, which separated EEG into predefined frequency bands, the convolutional recurrent neural network can automatically learn convolutional filters to extract the weighted combinations of EEG channels and then use the recurrent module to process sequential information for accurate classification and subtype discovery.

This disclosure describes a state-of-the-art deep learning method which employs a multi-scale convolutional recurrent neural network, MCRNN, to leverage the spatiotemporal information of rsEEG for multi-disease classification and subtype discovery. By initially learning profiles from three major psychiatric disorders with comorbidity comprehensively, the model can effectively inhibit the confounds (e.g. age, gender, scanners) by mapping the EEG features into the psychiatric-specific subspace, making unsupervised clustering practical for subtype discovery. In the example presented below, the model achieved and appreciable average 4-class classification accuracy (52.4%) when applied to a multiple psychiatric disorder dataset consisting of BP, MDD, SZ, and normal controls (NC). The knowledge learned by the deep learning model may be further used for clustering MDD patients. By applying the model to an independent validation dataset that consists of MDD, the MDD subtype which is sensitive to the TMS treatment was discovered, showing great promise in promoting the understanding of major psychiatric disorders and facilitating the individualized treatment of depression.

An overview of the methodis outlined in. A database of EEG measurements on human subjects with known diagnoses was generated and later accessed at step. In addition, also at step, other known information about the patients may be included, such as the patients' sex and age and also their response to TMS-treatment of MDD. The EEG data is accessed and pre-processed at step. The EEG data is divided into training and testing groups, and then used to train a deep learning classifier at step. In addition, at step, EEG measurements from human patients classed as normal (NC) may also be used in addition to the data from subjects with a known mental illness diagnosis. Both sets of data (normal controls and patients diagnosed with a known mental illness) may be used to train the deep learning classifier at step. Once the classifier has been trained and tested, it may be used to help develop an additional model to predict the TMS-treatment outcome at step. This additional model may use EEG scans of patients with an unknown diagnosis at step. These scans are pre-processed also at stepand then fed into the original trained model at step. A diagnosis or classification for a mental disorder class (e.g., MDD, BP, SZ, or NC) for each patient's EEG scans may be made at step. After these EEG scans have been classified, a prediction as to the TMS-treatment response may be made for patients classified in the MDD group at step.

As shown in, two datasets may be used: a first datasetof rsEEG samplings from patients with multiple psychiatric disorders and a second datasetof samples of patients with major depressive disorder who were treated with TMS. In an example, the first dataset of 4-class rsEEG samplings were preprocessedand then sent to a multi-scale recurrent neural network (MCRNN) for model optimization. After the preprocessing, EEG samples in the first datasetwere used for optimizing the parameters of MCRNN. A leave-one-out strategy may be used for evaluating the model performance. The severity continuum of multiple psychiatric disorders was visualized. The model interpretation was applied for discovering the most discriminative EEG channels. The second datasetwas sent to the optimized MCRNN for feature extraction, and the extracted featureswere further used for unsupervised MDD subtype discovery. Two clustered subtypes are compared in the aspects of TMS response prediction, HDRS, functional connectivity, and neural plasticity.

The leave-one-out cross-validation strategy was used for evaluating the classification performance. The severity continuum of psychiatric disorders was visualized based on the high-level feature learning by MCRNN. To identify the most discriminative EEG channels, an occlusion strategy was applied for explaining the discriminative power of EEG channels. To identify the TMS-response subtype, the optimized MCRNN model was further applied to an independent validation MDD dataset (the second dataset: for example, 25 subjects with TMS treatment), and two MDD subtypes were discovered by unsupervised clusteringof the features extracted using MCRNN, as shown in. The two MDD subtypes were compared in TMS treatment response, Hamilton Depression Rating Scale (HDRS), functional connectivity, and neural plasticity.

Some details of the MCRNNare shown in. EEG voltage signals were measured as a function of time. Various filters were applied to these signals in a set of multi-scale convolutions. The model has two main modules: multi-scale 1D convolutional layers as filters to map the preprocessed EEG sampling into various feature spaces and a gated recurrent unit (GRU) for aggregating sequential information. The fully connected layer at the end will classify the input set of EEG voltage signals as belonging to one of the four categories: BP, MDD, NC, or SZ.

In the example shown in, the MCRNNmay consist of multiple 1D convolutional (Conv1D) filters with different scales, one concatenation layer, one max-pooling layer, a gated recurrent unit (GRU), and an averaged layer for integrating the spatiotemporal information for classification. The preprocessed EEG signals were fed into the proposed MCRNN model for parameter optimization. After optimizing the parameters, the model was saved for performance evaluation. The detailed architecture and mechanisms of the MCRNN model are as follows: The multi-scale 1D convolutional layer expands upon simple convolutional layers by including multiple filters of varying sizes in each Conv1D layer. The filter lengths used in the Conv1D are drawn from a logarithmic instead of a linear scale, leading to exponentially varying filter lengths (2, 4, and 8). Therefore, the dimensions of 3 different scales of convolutional filters are 64 (EEG channels)×2 (filter length)×32 (number of filters), 64×4×32, and 64×8×32. A concatenation layer then concatenates the incoming features among the depth axis, resulting in feature maps whose sizes are 250 (time points)×96 (feature dimensions). Whereafter, a max-pooling layer performs a down-sampling operation along the time dimensions with filter size 3, resulting in features whose size is 83 (time points)×96 (feature dimensions). The downsampled features are the input of the subsequent GRU layers. As for the GRU layer, the size of the GRU's hidden state was set to 32. The GRU layer can extract the sequential information and hidden states of the EEG signals. The extracted hidden states are then sent to the average-pooling layer to combine all GRU steps. The fully connected layer and SoftMax are then applied to get the final prediction results. More details of the model implementation are presented below.

The MCRNN model was trained by minimizing the cross-entropy loss using the Adam optimizer. The training batch size was set to 512. The learning rate started from 0.001 and decayed after each epoch with a decay rate of 10. To improve a generalization of the performance of the model and to overcome overfitting, dropout (dropout rate: 0.5 in the convolutional module, 0.3 in the GRU module, 0.5 in the fully connected layer) and Lregularization (GRU kernel regular l=10, l=10) were also applied for regulating the model parameters. The training process was stopped when the validation loss stopped decreasing for 50 epochs or when the maximum epochs (1000 epochs) had been executed. The intermediate model which achieved the highest accuracy on the validation dataset was reserved for testing. The proposed models were implemented on TensorFlow platform.

Each epoch is represented with a T×D matrix, where T is the length of samplings and D is the amount of EEG channels. To quantify the classification contribution of the dchannel, the samplings of dchannel were replaced with its averaged value while keeping other channels' samplings as they were. This replacement is equivalent to eliminating the contribution of dchannel. All the testing samples were processed in the same way and subsequently fed to the trained MCRNN model. The classification performance of the trained model which is fed with reduced features would decrease compared to the performance using all features. The channels which maximize the decrease of the classification performance are further selected as the most discriminative channels.

In an example, the first datasetconsisted of 185 subjects (51 BP, 46 MDD, 41 SZ, 47 NC=normal control) was used for training the MCRNN model. The leave-one-out strategy was applied for evaluating the classification performance. The average accuracy of the 4-class classification achieved 52.4% (). Table 1 shows the number of patients and corresponds to. The confusion matrix shows that MDD and BP exhibit more overlap than other psychiatric disorders. Permutation tests of overall accuracy for the classifier were conducted and the results shown in. For each permutation run, the labels of subjects were randomly shuffled. The result shows the null distributions of four-class classification accuracy from the empirical tests peaks at 0.28±0.06 (95% confidence intervals), whereas the MCRNN is at approximately 0.51, indicating that it is highly improbable that the MCRNN results are the result of chance level. To visualize the severity continuum of various mental disorders, a t-distributed stochastic neighbor embedding (t-SNE) was applied to the 32-dimensional features extracted from the second-last layer of MCRNN.shows the “spectrum” of multiple psychiatric disorders. The severity spectrum shows that MDD and BP overlap, coinciding with the confusion matrix results.

All epochs of 185 subjects in the first datasetwere used for optimizing the parameters of MCRNN. After optimizing the trained model, the parameters of the trained model were saved. All EEG epochs without removing any channel were sent to the MCRNN for obtaining the benchmark classification performance. Afterward, the EEG epochs with removed information of one specific channel were fed to the model for obtaining the decline of classification accuracy repeatedly. Then, the decline of accuracy when removing a specific channel was recorded and sorted. In this way, the contribution of each channel was calculated. As shown in, the most discriminative channels are located in Occipital (O1, O2, PO4), Parietal (P3, P7, P8), and Frontal (Fp1, F8). These most discriminative channels are highlighted with thicker borders in the figure. Details of the EEG electrodes can be found elsewhere in this disclosure.

The understanding of psychiatric disorders with neurobiological heterogeneity could benefit from the identification of subtypes discovered using state-of-the-art neuroimaging techniques. Deep learning, which can learn a comprehensible task-specific projection spectrum from the high-dimensional samples, is promising in subtype discovery. In this example, after training the MCRNN model on the first dataset, the optimized MCRNN learned the mapping to project the EEG epochs to a subspace in which the psychiatric disorders are maximally separated. Afterwards, when the EEG epochs of the independent validation dataset (the second dataset) were sent to the optimized MCRNN, the epochs were mapped to the psychiatric-related subspace in which the non-psychiatric confounds (e.g., age, gender, scanners) are maximally inhibited, making the extracted features from MCRNN robust for subsequent subtype discovery. The t-SNE dimension reduction and visualization algorithm was subsequently applied to the extracted features, though other clustering algorithms may also be used. Two distinct groups are shown after this method was applied (). The two subtypes are further compared in the following four aspects: (1) TMS treatment response; (2) Hamilton depression rating scale; (3) functional connectivity patterns; (4) neural plasticity.

It is very challenging to predict the TMS response with the original EEG signals because of their low signal-to-noise ratios. Therefore, the high-level features were extracted first from the trained MCRNN model and then these features were used to train a support vector machine (SVM) model to predict the TMS-treatment response.

During the training procedure, the MCRNN was firstly optimized using the training dataset. After training, the parameters of the model were saved (“MCRNN_trained”). The high-level features of the training dataset were extracted using the trained model. An SVM was then used to predict the TMS-response using these high-level features. A leave-one-out cross-validation (LOOCV) strategy was employed to evaluate the SVM prediction performance. Since each subject has 260˜300 epochs, during the training procedure, the label of the epoch was set to the same label as the subject when training the classifiers. The predicted results of epochs were used to vote and obtain the final label of the subjects when validating. The hyperparameters are also saved (“SVM_trained”) to the model.

When the model was applied to a new undiagnosed patient for diagnosis (classification) and TMS-treatment response prediction, the preprocessed EEG epochs are firstly sent to the “MCRNN_trained” model for high-level feature extraction. Afterwards, these extracted high-level features are then sent to the “SVM_trained” model to get the TMS-treatment response prediction.

As shown inand Table 2, the clustered subtypes are highly correlated to the TMS treatment response, indicating that the TMS response may be predicted before treatment using rsEEG. To further quantify the ability of extracted rsEEG features in predicting the TMS treatment response, a support vector machine (SVM) and leave-one-out cross-validation (LOOCV) strategy was applied to evaluate the prediction performance using MATLAB. The kernel function of the SVM was “radial basis function” and other SVM parameters were set as default. Since each subject had 260˜300 epochs, during the SVM training procedure, the label of each epoch was set the same as its corresponding subject. During testing, the predicted results of epochs were used to vote for obtaining the final label of the subject. The classification result is shown inand Table 2 and the receiver operating characteristic (ROC) is shown in. The TMS treatment response could be predicted using SVM based on high-level rsEEG features extracted using MCRNN with a mean accuracy of 84%. The area under the curve (AUC) is 0.82, which also demonstrates that this method is considered highly accurate.

Before and after the TMS treatment, the Hamilton Depression Rating Scale (HDRS) of each subject was collected. Five factors (F1: Psychic depression; F2: Loss of motivated behavior; F3: Psychosis; F4: Anxiety; and F5: Sleep disturbance) were further calculated according to Milak et al.'s factor analysis results. (See Milak et al., “Neuroanatomic Correlates of Psychopathologic Components of Major Depressive Disorder”62, 397-408, (2005) for details). The total HDRS reduction, F3 (Psychosis), and F4 (Anxiety) of two subtypes are visualized inand exhibit a clear gradient. As shown in, the total HDRS reduction and the change of the five factors pre- and post-TMS treatment are quantitatively compared between the two subtypes. Compared to subtype 1, subtype 2 has a significant reduction in HDRS total score, indicating a positive response to the TMS treatment. Subtype 1 shows a significantly higher value than subtype 2 in F3 (psychosis) factor (p<0.05, two-sample t-test) before TMS treatment, indicating a potential biomarker for TMS treatment outcome prediction. Besides, subtype 2 exhibits significant (p<0.05, two-sample t-test) relief of symptoms in F1 (psychiatric depression), F2 (loss of motivated behavior), and F5 (anxiety) after TMS treatment.

The coherence functional connectivity of the alpha-band (10 Hz) was calculated using the Fieldtrip toolbox. For source reconstruction, the precomputed MNI-standard Desikan-Killiany atlas was loaded (details shown in Table 3), a boundary element method (BEM) head model, and the source model. The whole-brain connectivity patterns based on the imaginary part of coherency are shown in. The t-test statistic results show the connection between the right rostral middle frontal and left superior parietal shows the most significant difference in the two MDD subtypes (p<0.005, two-sample t-test). Functional connectivity of subtype 1 is shown in, subtype 2 is shown inand t-test results are depicted in. The connection between ‘rostral middle frontal R’ (frontal) and ‘superior parietal L’ (parietal) shows significant differences between two subtypes (p<0.005, two-sample t-test) as are indicated by the arrows in the.show details of the diagonal of the matrix for the number region of interest based on the Desikan-Killiany atlas shown in Table 3. The diagonals for all three matrices show the same numbers, so the diagonal only foris shown here, but the corresponding diagonals forare the same.

As shown in, the subjects in two MDD subtypes behave differently in event related potentials (ERPs) when conducting arrow version flanker tasks. Some details of the flanker task are shown inand more details supplied elsewhere in this disclosure. Compared to subtype 1, the subtype 2 patients show higher P300 amplitude on electrode Fp1 [F(1,23)=15.12; p=0.001], P3 [F(1,23)=5.806; p=0.026], P7 [F(1,23)=4.643; p=0.044], O1 [F(1,23)=5.669; p=0.027] and O2[F(1,23)=4.765; p=0.037], indicating that the subtype 2 had higher neural activity during the task.

Multiple psychiatric disorders dataset (the first dataset). According to the Diagnostic and Statistical Manual of Mental Disorders (DSM-5), a total of 185 participants (all Chinese Han people) (BP=51, MDD=46, SZ=47, NC=41) were diagnosed after the semi-structured clinical interview by experienced psychiatrists (shown in Table 4). Patients in the first datasetmeet DSM-5 criteria for bipolar disorder, major depressive disorder, or schizophrenia. The participants were within the 14-40 age range, right-handed. Participants were excluded from enrollment if they had a currently active substance-use disorder, brain injury, a history of seizures, unstable medical condition, current pregnancy, prior electroconvulsive therapy. Healthy controls were recruited from the local community. The healthy controls also received semi-structured clinical interviews to exclude any current or lifetime evidence of psychiatric disorder. All participants in this study signed written informed consent, which was approved by the ethics committee of the institutional review board of Second Affiliated Hospital, Zhejiang University School of Medicine. Details of the protocols for EEG are shown in elsewhere in this disclosure.

MDD with TMS treatment dataset (second dataset): All subjects (shown in Table 5) were diagnosed as MDD by experienced psychiatrists after a semi-structured clinical interview and DSM-5. The subjects were within the 16-48 age range, right-handed who were screened for ethical clearance, with only Chinese Han people. Specific clinical symptoms were evaluated using the Hamilton depression rating scale. The EEG signals were recorded using the same equipment as the first dataset. In addition, all the MDD patients participated the event-related potential (ERP) task (the Arrow version flanker task, more details of the task can be found below). Hamilton depression rating scales were rated before and after a complete course of TMS treatment. The TMS protocol involved 5 consecutive days of two intermittent theta-burst stimulation (iTBS) sessions per day (1,800 pulses per session, the interval between each session is above 4 hours) delivered to the region of the left dorsolateral prefrontal cortex (left DLPFC). The iTBS was performed using a MagPro X100 stimulator (Mag Venture, Georgia, USA) with a figure-of-eight coil, 70 mm diameter. The stimulus intensity is 110%˜120% of the resting motor threshold.

Details of the flanker task.shows some details of the flanker task. In each trial, participants viewed five white arrows presented for 500 ms against a black background as shown in. The participants were asked to respond as quickly and as accurately as possible to indicate the direction of the middle arrow by pressing the left or right mouse button. Participants had up to 3000 ms from the onset of the stimulus to respond. 200 of the trials were congruentin that all the arrows were pointing in the same direction (←←←←← or →→→→→) (as in part A of), whereas the other 40 trials were incongruent 3606 (e.g., ←←→←← or →→←→→) (as in Part B of). The trial order was random for each participant and the inter-trial interval varied from 2000 to 3000 ms, during which time participants viewed a white fixation crosscentered on a black background. Participants received short breaks every 80 trials throughout the task.

By training the MCRNN on the first datasetin a supervised way, the model has learned the way to nonlinearly map the EEG features from its original feature space to a subspace in which the differences among psychiatric disorders are more distinctive. To make full use of the advantage of MCRNN, the trained MCRNN was applied to an independent MDD dataset for subtype discovery. The procedures are as follows: After optimizing the MCRNN model using The first dataset, the model parameters of the MCRNN were saved. The EEG signals of each subject in The second datasetwere first preprocessed using the same procedure as The first dataset, then the epochs were sent to the trained MCRNN model for extracting the hidden state feature. As a result, for each EEG epoch in The second dataset, a vector representation consisting of 32 elements was obtained. The epochs were pooled together for t-SNE clustering and subtype discovery.

Due to the intrinsic heterogeneity and strong comorbidities in neurobiological abnormalities within current clinical definitions for diagnosis, discovering objective psychiatric biomarkers from neuroimaging is vital for understanding pathophysiology and improving treatment. In this disclosure, a new deep learning-based model, MCRNN, for identifying multiple psychiatric disorders using rsEEG is described. An average accuracy of 52.4% was achieved in the 4-class classification task, which is significantly above that expected by chance alone (25%). The spectrum continuum of psychiatric disorders was then visualized using t-SNE based on the extracted rsEEG features. Subsequently, the MCRNN optimized on the multiple psychiatric disorder dataset was applied to an independent MDD dataset, leading to the discovery of an MDD subtype which is sensitive to the TMS treatment. The various subtypes also exhibited significant differences in Hamilton depression rating scores, functional connectivity patterns, and neural plasticity.

It is challenging to quantify diagnosis based on the clinical symptoms. The current classification diagnostic criteria (ICD and DSM) are based on the evaluation of clinical symptoms. The same group of similar symptoms may be caused by completely different biological processes. The present disclosure classified mental diseases based on objective neurophysiological markers. As a neural activity recording tool with high time resolution, rsEEG can classify the neural activities of different mental diseases at the systematic level. Combined with the state-of-the-art deep learning methods, EEG exhibits great prospects in facilitating the accuracy of clinical diagnosis by revealing the mechanisms of psychiatric disorders. Compared to conventional time-frequency-based analysis which manually extracts different frequency bands of rsEEG to generate functional connectivity features, the convolutional module of the MCRNN can automatically learn spatial filters to map the original EEG into subspaces. The recurrent module subsequently integrates the sequential information for accurate classification. The 4-class results demonstrate that the MCRNN can efficiently capture the discriminative neural activity patterns among psychiatric mental disorders (e.g., MDD, BP, SZ). By analyzing the 4-class confusion matrix and the psychiatric disorder visualization, it was found that BP and MDD have more bilateral overlaps than others. This phenomenon occurs because BP and MDD have many overlaps in clinical symptoms, and the core symptom of MDD can also be found in the depressive or mixed states of BP. With this deep learning model, the specific stages of the disease (such as prodromal stage, initial stage, and chronic disease stage) can be quantified, visualized, and compared. In addition, the treatment response can be accurately predicted, which benefits patients with accurate individualized intervention. In this example, the DSM label was used as the ground truth to train the 4-class deep learning model, this procedure introduced some biases to the model and would be overcome using an unsupervised learning approach in the future.